[Date Index] [Thread Index] [Author Index]
Re: weibull plot on weibull scaled paper
On 2/24/11 at 6:21 AM, darreng at wolfram.com (Darren Glosemeyer) wrote: >On 2/23/2011 4:24 AM, Bill Rowe wrote: >>But if I enter each line by hand into individual cells, I don't see >>the error. So, there appears to be a bug somewhere. That is, the >>various ways I've described above of doing the computation should >>all behave identically. >ProbabilityScalePlot and its friends call the internal parameter >estimation code for FindDistributionParameters/EstimatedDistribution >to get the estimates. The reason for the message is that the >equations in the estimation code for WeibullDistribution were less >numerically stable than they could have been and so convergence >fails more often than it should when parameter values are somewhat >large. This has been fixed for the next release. >I think the fact that the example worked without error when you >typed the inputs in separate cells is a red herring. The example >will converge for some data sets and not for others. I suspect the >random values you got in the case that worked just happened to give >a convergent result while the random values you'd gotten in the >other attempts did not. Your explanation makes sense and I should have thought about the random number generator. But I was getting such consistent results, this just didn't occur to me. Possibly part of the reason I had never seen this in the past when I've used random Weibull deviates I've been interested in smaller shape factors (in the range of 0.2 to ~5). That also might change the likelihood of getting values leading to the error. Additionally, when I was attempting to isolate the problem, I was actually quitting the kernel and restarting it so that each test started with a fresh session. This too, might have had an affect on how likely it was to get a set of values that lead to the error message.